yinita/gigpo-sft-w3c20-sync1-0726-step30

gigpo-sft-w3c20-sync1-0726 — step 30

LoRA adapter from GiGPO run gigpo_sft_w3c20_sync1_50_0726_0459 at step 30.

Run config

• Base: yinita/ps4mas-sft-x5-single-ep3
• Judge: Claude Sonnet 4.6 (Bedrock)
• Reward: phase2-only (acq_w=3, comp_w=20), reward_phase1_steps=0
• sync-every=1, force_acquire_frac=0, max_steps=50
• WandB: https://wandb.ai/yinita/PS4MAS_GiGPO/runs/cgtq25cv

Case study @ step 30 (n=10, Sonnet)

• mean_composite: 2.532
• acquire_rate: 0.20

Series

repo	step	CS composite	CS acq
yinita/gigpo-sft-w3c20-sync1-0726-step10	10	3.066	0.50
yinita/gigpo-sft-w3c20-sync1-0726-step20	20	2.732	0.40
yinita/gigpo-sft-w3c20-sync1-0726-step30	30	2.532	0.20
yinita/gigpo-sft-w3c20-sync1-0726-step40	40	2.600	0.10
yinita/gigpo-sft-w3c20-sync1-0726-step50	50	2.499	0.10

Best CS in this series: step10. Adapter-only (no optimizer). Load as LoRA on the SFT base (e.g. vLLM --enable-lora).